Pritesh Kumar

Built new algorithms from scratch for new products/apps like news aggregators, etc by: Fine-tuning sentence transformers like BERT, MPNet, etc for downstream tasks like text classification, Semantic Textual Similarity(STS) and NER-extraction, attaining an accuracy of 95%. Created unique logic for sections of news app like ‘For You’, ‘Trending’, etc, using ML-models like BERT. Developed a novel logic for batching process using ML-modeling. Memory and Speed optimization of the algorithms: reducing memory-usage and speed by 80% using multi-threading, etc. Did API-integration of text summarization models like ‘t5-base’, Facebook’s BART, Pegasus, etc for various in-house applications. Git-integration of the pipelines built. Created a new cleaner package module in “pypi.org” for easier integration to the pipelines. Generated relevant visualizations to assess the performance of various ML-models for multiple tasks using Seaborn and Matplotlib. Integrated database with the pipeline and the API for efficient running of the pipelines. Performed Data Engineering: setting up MongoDB and BigQuery, and also performed database optimization on it. Deployed ML-models and their pipelines on the server. User Recommendation System: Scraping data using Beautiful-Soup and Selenium to be used for user-tagging. NER-extraction from text using Spacy to extract important entities. User-tagging of NER-entities and news articles. Visualizing user app-interaction for better understanding of the recommendation system and better planning for user-targeting using Seaborn, Matplotlib, etc, leading to an increase in user retention-rate by 20%. Computer Vision: Designed and created pipelines for user-recommendation system using various huggingface models for: Video to text conversion: Saleforce’s blip-2 model, Audio to text conversion, Image to text conversion. For all the above tasks created a very useful unique strategy to translate video/image/audio elements into text for user-recommendation system